CN109343505A - Gear method for predicting residual useful life based on shot and long term memory network - Google Patents
Gear method for predicting residual useful life based on shot and long term memory network Download PDFInfo
- Publication number
- CN109343505A CN109343505A CN201811096152.0A CN201811096152A CN109343505A CN 109343505 A CN109343505 A CN 109343505A CN 201811096152 A CN201811096152 A CN 201811096152A CN 109343505 A CN109343505 A CN 109343505A
- Authority
- CN
- China
- Prior art keywords
- gear
- lstm
- input
- cell
- gradient
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B23/00—Testing or monitoring of control systems or parts thereof
- G05B23/02—Electric testing or monitoring
- G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
- G05B23/0259—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterized by the response to fault detection
- G05B23/0283—Predictive maintenance, e.g. involving the monitoring of a system and, based on the monitoring results, taking decisions on the maintenance schedule of the monitored system; Estimating remaining useful life [RUL]
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
- G01M13/02—Gearings; Transmission mechanisms
- G01M13/021—Gearings
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
- G01M13/02—Gearings; Transmission mechanisms
- G01M13/028—Acoustic or vibration analysis
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Automation & Control Theory (AREA)
- Acoustics & Sound (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention is based on the gear method for predicting residual useful life of shot and long term memory network to belong to big data and intelligence manufacture field, it solves the problems, such as to rely on phenomena such as being easy to appear gradient explosion or gradient dispersion with Recognition with Recurrent Neural Network when existing RNN algorithm can not solve long, using algorithm, specific step is as follows: 1, using sensor to power gear degeneration real-time monitoring;2, feature extraction is carried out to gear fatigue state, slump evaluations is carried out to gear wear degraded performance;3, the Real-time Monitoring Data prediction model of the tooth bending fatigue based on LSTM is established;4, the parameter based on LSTM prediction model is preferred;5, gear predicting residual useful life is carried out according to gear condition estimation and known gear distress threshold value and provides foundation it is an advantage of the invention that the accuracy of gear degenerate state and real-time remaining life can be predicted effectively for gear prospective maintenance.
Description
Technical field
The invention belongs to big datas and intelligence manufacture field, more particularly to a kind of gear based on shot and long term memory network
Method for predicting residual useful life.
Background technique
Gear-box is the machinery field such as main mechanical part of cluster engine, wind power system, gear drive be even more along with
Entire steel industry.With the expansion of modern machinery and equipment collection on a large scale, the continuous improvement of precision, long-term operation middle gear very
It is easy to appear that attrition causes the gear teeth thinning or stress is concentrated, tooth root decrease of fatigue strength reduces gear life in turn, weighs
Carry lower even generation break of gear tooth.The health status of monitoring gear is to ensure mechanical equipment reliability service important means.
Gear is generally enclosed in gear-box, the voice signal and vibration signal real-time monitoring received by sensing equipment
Gear health state, but the characteristic quantity for generating the vibration signal of failure has extremely complex non-linear relation and enchancement factor.
The data that characterization equipment degenerate state is obtained by monitoring carry out fault identification and pre- by data processing and feature extraction again
It surveys.Failure is identified and predicted by model and method, realizes the life prediction of mechanical equipment.
In recent years, with the continuous development of artificial intelligence technology, deep learning is also had received widespread attention.Deep learning
Essence is the neural network model for having many hidden layers by the training data building of magnanimity, the feature by sample in former space
Expression transforms to a new feature space, successively extracts potential mapping relations in input and output sample.Recognition with Recurrent Neural Network
(Recurrent NeuralNetworks, RNN) is chiefly used in the processing of sequence data, can remember to the sequence inputting of front
And in the calculating after being applied to.RNN uses Back-propagation Through Time after completing layer-by-layer pre-training
(BPTT) residual error of accumulation is passed back and completes parameter regulation by algorithm.Dependence when BPTT can not solve long is found in practical application
Problem, i.e., current output is related with one section of very long sequence of front, is easy to bring gradient disappearance or gradient explosion issues.Length
Phase remembers (long short-term memory, LSTM) network model and passes through the gate of more parameters and control information flow
Cellular system can avoid the problems such as gradient is disappeared with gradient explosion, so that timing information prediction is more accurate.
Summary of the invention
The present invention overcomes the shortcomings of the prior art, what is solved existing BPTT algorithm and rely on when can not solve long is asked
Topic and Recognition with Recurrent Neural Network are easy to appear phenomena such as gradient explosion or gradient dispersion, it is desirable to provide one kind is based on modified length
The gear method for predicting residual useful life of phase memory network, the present invention include three layers of (input layer, hidden layer and output layer) network knot
The detailed design of structure and the realization algorithm of network training and neural network forecast etc..On this basis, mould is selected using random search
Type hyper parameter is added that momentum considers and builds from network structure, improved model is accurate in the settings such as learning rate, window size
Property, it prevents from falling into local optimum, improves model prediction accuracy.
In order to solve the above technical problems, the technical solution adopted by the present invention are as follows: based on modified shot and long term memory network
Gear method for predicting residual useful life, including following implementation steps:
Step 1 installs sensor in gear-box, obtains the Real-time Monitoring Data of tooth bending fatigue;
Acceleration transducer is mounted on the bearing block position of main examination case, the mounting temperature sensor in gear-box, in main examination
Noise transducer is installed in the surface of case,
Step 2 carries out feature extraction to the fatigue state of gear, carries out slump evaluations to gear wear degraded performance;
Gear wear condition is changed using square amplitude and carries out performance slump evaluations, for each sampling time Δ t long
In degree, the square amplitude of the time series of discrete random signal be may be expressed as:
Δ t is using the time in formula;N=Fs× Δ t, FsFor sample frequency, n is sampling number, and ∑ indicates summation, i ∈
(1,2,3....n), xiIt (t) is sampled value;
Step 3, the Real-time Monitoring Data prediction model for establishing the tooth bending fatigue based on LSTM:
The present invention constructs the general frame of LSTM prediction model as shown in Figure 1, including input layer, hidden layer, output layer, net
Five functional modules of network training and neural network forecast.Input layer is responsible for carrying out the Real-time Monitoring Data of tooth bending fatigue preliminary
To meet network inputs requirement, hidden layer builds single layer Recognition with Recurrent Neural Network, output layer using the LSTM cell that Fig. 2 is indicated for processing
Prediction result is provided, network training uses building from network structure, optimizes, join to LSTM on learning rate, window setting
Nesterov momentum is added when number amendment, prevents from falling into local optimum, using the excellent of RMSProp algorithm correction model adaptive rate
Change method, neural network forecast uses the method prediction point-by point of iteration, and then improves the accuracy of network model prediction.
LSTM recirculating network as shown in Figure 2 (replaces general circulation nerve at itself " cell " other than RNN formula circulation
The hidden unit of network) inside circulation.Each unit phase input and output, but also have more parameters and control information flow
Door control unit system, i.e. state cell(moment t, cell i), external input door (external input gate) forget door
(forget gate), out gate (output gate).Forget door to be responsible for removing the unessential information of LSTM study from " cell ",
These information will be removed by door control unit operation.Forget door and takes two input ht-1And xt。ht-1For the state of previous unit
Value, xtFor the value of the t moment of list entries x.Add bias term again to input Sigmoid function.Sigmoid function value range
For [0,1], if value is zero, forgetting door allows for location mode and forgets the information.Corresponding propagated forward algorithm is as follows:
Input gate:
Forget door:
Cell:
Out gate:
Cell output:
Backpropagation is calculated as shown in (8)-(14):
Cell output:
Out gate:
State:
Cell:
Forget door:
Input gate:
Wherein ∑ indicates summation,It is current input vector, subscript l, φ, c, ω, which is respectively indicated, to be indicated input gate, lose
Forget the relevant parameter of door, cell state unit, out gate.wil、wiφ、wicAnd wiwRespectively indicate from be input to input gate, forget
The weight of door, cell state unit, out gate, whl、whφ、whcAnd whwIt respectively indicates from hidden layer to input gate, forget door, cell
The weight of state cell, out gate, wcl、wcφAnd wcwIt respectively indicates from cell state to input gate, the power of forgetting door and out gate
Weight, atIndicate the input of t moment, btIndicate the output of t moment,Indicate the state of t moment cell, g and h respectively indicate cell
Input and output activation primitive, I indicate input layer neuron number, K indicate output layer neuron number, H indicate it is hidden
The number of cells of layer, σ is sigmoid function.Indicate the cell output of t moment,Indicate the state of t moment, ζ indicates loss
Function, δtIndicate gradient error, wckAnd wcgIt respectively indicates from cell state to output layer and the weight of hidden layer,
Network training
First in input layer, the Real-time Monitoring Data that we define tooth bending fatigue is F0={ f1,f2,…,fn,
The training set and test set then divided can be expressed as Ftr={ f1,f2,…,fmAnd Fte={ fm+1,fm+2,…,fn, meet about
Beam condition m < n and m, n ∈.Then to the element f in training settIt is standardized, is standardized using classical z-score public
Formula (mean value 0, standard deviation 1, be expressed as zscore), the training set after standardization can indicate are as follows:
Ftr'={ f1',f2',…,fm'} (15)
Meet:
1≤t≤m,t∈ (17)
The characteristics of in order to adapt to hidden layer input, we apply the method f ' of data segmentationtrIt is handled, sets segmentation window
Mouth length value is L, then the mode input after dividing are as follows:
X={ X1,X2,…,XL} (18)
Xp={ fp',fp+1',…,fm-L+p-1'} (19)
Meet:
1≤p≤L;p,L∈N (20)
Corresponding output are as follows:
Y={ Y1,Y2,…YL} (21)
Yp={ fp+1',fp+2',…,fm-L+p'} (22)
Next, X is inputted hidden layer by us.It will be seen from figure 1 that hidden layer includes L by the connection of front and back moment
Isomorphism LSTM cell, the output after hidden layer can indicate are as follows:
P={ P1,P2,…,PL} (23)
Pp=LSTMforward(Xp,Cp-1,Hp-1) (24)
Wherein, Cp-1And Hp-1The state and output of respectively previous LSTM cell, LSTMforwardBefore indicating LSTM cell
To calculation method (formula (2)-(7)).Cell state vector magnitude is set as Sstate, then Cp-1And Hp-12 vector magnitudes are
Sstate.As can be seen that hidden layer P, mode input X and theoretical output Y are the two-dimensional arrays that dimension is (m-L, L).It selects equal
For square error as error calculation formula, the loss function of training process can be with is defined as:
The minimum optimization aim of loss function is set, the train epochs steps of netinit is given, using RMSProp
Change learning rate, Nesterov introduces momentum and changes gradient, changes update mode in terms of two, and then obtain final hidden layer
Network.
The present invention optimizes loss function using stochastic gradient descent (Stochasitc Gradient Descent, SGD).
Gradient descent method will minimize loss on whole training datas, can be very when sample size is very big or the number of iterations increases
Consume computing resource.Stochastic gradient descent generates distribution according to data and extracts m small lot (independent identically distributed) sample, each
Secondary iteration only calculates the loss of a sample, then traverses all samples again, completes the calculating of a wheel.By calculating their ladder
Mean value is spent, the unbiased esti-mator of available gradient is expressed as follows
Wherein g is gradient, and θ is the initial parameter of LSTM model forward calculation,It is gradient element, table
ShowLocal derviation is asked to θ, using the sample from small lot m, stochastic gradient descent algorithm uses following gradient
Decline estimation
θ←θ-ηg (27)
Wherein η is learning rate, and L is the loss of each sample
L (x, y, θ)=- logp (y | x;θ) (28)
Wherein, logarithm is sought in log expression.
Crucial parameter is learning rate in SGD algorithm, can gradually change study over time in LSTM application
Rate.
Momentum parameter optimization
In deep learning, stochastic gradient descent is the optimization algorithm being in daily use, and optimization process is very slow sometimes, therefore
Momentum optimization algorithm is added in the present invention to be intended to accelerate to learn, especially to the gradient of plus noise.Momentum is in physically table
Speed is shown as multiplied by quality, in momentum learning algorithm, it is assumed that be unit mass, therefore velocity vector v can regard the dynamic of particle as
Amount.The inertia for simulating object, gradient is added up, that is, increases stability to a certain extent when update, thus accelerate to learn,
Inhibit concussion, there is the ability for jumping out local optimum, be expressed as
The contribution decaying of gradient, is set as 0.5,0.9,0.99 before hyper parameter α ∈ [0,1] is determined.If Error function
Gradient g is always observed, then it can ceaselessly accelerate on-g, until reaching final speed.Wherein step sizes are
Wherein momentum hyper parameter isWhen α=0.9,10 times of gradient descent algorithms of maximum speed are corresponded to.Make herein
With the stochastic gradient descent of Nesterov momentum, error is calculated after parameter Estimation is complete, is implemented as follows shown:
Assuming that learning rate η, momentum parameter α, initial rate v
It is iterated process every time:
1. extracting the small lot sample of { x (1), x (2) ... ..x (m) } from training set, wherein xtCorresponding output yt,
2. calculate gradient and Loss, v and θ are updated with this:
Calculate gradient:
It is interim to update:
θ←θ+v
Difference between Nesterov momentum and standard momentum is embodied in gradiometer and counts in, gradiometer in Nesterov momentum
It calculates after applying present speed, it is optimal in conjunction with original gradient searching, it may be interpreted as being added to one into standard momentum method
Correction factor.
Variable learning rate arithmetic
Learning rate has a significant impact the performance of model, determines that parameter is moved to the speed of optimal value.Amplitude is excessive, meeting
Cause parameter that may cross optimal value;Amplitude is too small, easily causes operation redundancy, and long-time operation can not restrain.The present invention adopts
It with the RMSProp algorithm adaptively changing learning rate of Nesterov momentum, acquires optimal value and obtains while improving convergence rate, update
It is as follows:
Assuming that learning rate η, momentum parameter α, initial rate v, rate of decay ρ initialize cumulant variable γ=0
It is iterated process every time:
1. extracting the small lot sample of { x (1), x (2) ... ..x (m) } from training set, wherein xtCorresponding output yt,
2. calculating gradient and Loss, γ is updated, calculates and updates in conjunction with gradient:
Calculate gradient:
It is interim to update:
θ←θ+v
RMSProp changes learning rate, and Nesterov introduces momentum and changes gradient, changes update mode in terms of two.Every time
After iteration, the value of loss function is checked using model parameter.If being less than last iteration, increase learning rate;Conversely, reducing study
Rate.
Neural network forecast
We (are expressed as application enhancements type LSTM network) predicted.The method that prediction process uses iteration.
Firstly, the last line data of output Y are as follows:
Yf={ fm-L+1',fm-L+2',…,fm'} (31)
Y is inputtedOutput result can indicate are as follows:
Then the predicted value at m+1 moment is pm+1.Then, by YfRear L-1 data point and pm+1Merge into a new line number
According to:
Yf+1={ fm-L+2',fm-L+3',…,pm+1} (33)
By Yf+1InputThen the predicted value at m+2 moment is pm+2, and so on, obtained forecasting sequence are as follows:
Po={ pm+1,pm+2,…pn}(34)
Next, by P0It carries out z-score instead to standardize and (be expressed as de_zscore), obtain final with test
Collect FteCorresponding forecasting sequence are as follows:
Wherein: m+1≤k≤n, k ∈ N.
Similarly, available with training set F using every a line of X as mode inputtrCorresponding fitting sequence Ptr.Most
Afterwards, by calculating FtrAnd PtrAnd FteAnd PteDeviation quantitatively provide fitting and the precision of prediction of model.
Step 4, the parameter based on LSTM prediction model are preferred:
Deep learning algorithm all controls network parameter with many hyper parameters, and the optimization algorithm of automatic hyper parameter does not need to make
Determine the hyper parameter of learning algorithm.
It include many parameters in LSTM prediction model, wherein split window, state vector size is closed the most with learning rate
Key.Then, these hyper parameter cartesian products obtain one group of super parameter.Grid search uses a determining compensation, just
Then change each step on coefficient and learning rate and take a value, the present invention can be found that data are one-dimensional at certain by using random search
On variation it is more obvious, obtain apparent trend.
Random search programming is simple, using more convenient, can converge to the good value of hyper parameter faster.With grid search
Difference, random search do not need the value of discretization hyper parameter, and permission scans in a bigger set, without generating
Additional calculating cost.
The present invention uses random search, is that good hyper parameter can be quickly found because of it, the experiment not wasted, unlike
Grid search provides identical result to two different values of a hyper parameter (it is constant to give other hyper parameter values) sometimes.In net
In lattice search, other hyper parameters will possess identical value in this is tested twice, and in random search, they would generally have
Different values.Therefore, if the verifying collection error corresponding to the variation of the two values does not have significant difference, grid search does not have
It is necessary to repeat two experiments of equal value, and random search still can independently explore other hyper parameters twice.
Step 5 carries out gear predicting residual useful life according to gear condition estimation and known gear distress threshold value:
Data of the vibration data that acceleration transducer receives after feature extraction are as mode input, using improvement
Type LSTM model carries out gear predicting residual useful life.Pass through the degenerate state value predicted and known degenerate state fault threshold
The time for reaching fault threshold for the first time can be solved.Sampling number is converted to the time, absolute error percentage (Absolute is quoted
Percentage Error,APE)
Wherein, TRULFor the remaining life of model prediction,
Ta=T*-tp(38)
In formula, tpIndicate current time, TaFor the real surplus life-span of gear, T*Fault threshold is reached for the first time for gear
Time.
The invention has the advantages that
The present invention proposes a kind of real-time method for predicting residual useful life of gear, and the present invention is established for gear wear degenerative process
The real-time predicting residual useful life model of modified LSTM based on memory mechanism under a kind of deep learning frame, and from network structure
It builds, learning rate, window setting is upper optimizes to LSTM, and addition Nesterov momentum when parameters revision prevents from falling into part
It is optimal, using RMSProp algorithm correction model adaptive rate, and then improve the accuracy of network model prediction.
Detailed description of the invention
The present invention will be further described in detail with reference to the accompanying drawing.
Fig. 1 is the real-time method for predicting residual useful life flow chart of middle gear of the embodiment of the present invention;
Fig. 2 is LSTM structure chart in the embodiment of the present invention;
Fig. 3 is the square amplitude curve figure of characteristics extraction that No. 4 sensors receive data in the embodiment of the present invention;
Fig. 4 is to carry out gear predicting residual useful life effect picture using modified LSTM model in the embodiment of the present invention;Fig. 4
It (a) is 70 hours prediction effect figures, Fig. 4 (b) is 73.3 hours prediction effect figures, and Fig. 4 (c) is 75 hours prediction effect figures;
Fig. 5 is LSTM training error decline curve figure in the embodiment of the present invention;
Fig. 6 is grid search and random search figure in the embodiment of the present invention.
Specific embodiment
It is understandable to enable objects, features and advantages of the present invention to become apparent, with reference to the accompanying drawing to tool of the invention
Body embodiment is described in detail.
In the embodiment of the present invention, the real-time method for predicting residual useful life of gear, method flow diagram include as shown in Figure 1: with
Lower step:
Step 1 carries out fatigue test to gear, obtains the Real-time Monitoring Data that characterization gear is degenerated:
Gear fatigue life test uses power flow blocking test rack.The center of testing stand is away from for 150mm, motor
Revolving speed is 1200r/min.Test process is monitored cabinet vibration, oil temperature and noise etc..Use material for alloy in test
Steel, tooth face hardness are the hardened face gear of 58-61HRC, are surface-treated as carburizing and quenching.Engagement side is staggeredly overlapped using front and back sides
Formula.Main examination case module m=3, number of teeth z1=z2=50, pressure angle α=20 °, facewidth 29mm, the real work facewidth 13~
14mm;Accompanying examination case number of gear teeth is z3=z4=24, and pressure angle α=20 °, work facewidth 20mm.Gear mesh frequency has 2,
Main examination case gear is 1000Hz, and accompanying examination case gear is 480Hz.It tests lubricating oil and uses L-CKC320 Industrial Closed gear oil;
11 sensors are arranged in test altogether, and acceleration transducer is mounted on the bearing block position of main examination case, in gear-box
Mounting temperature sensor installs noise transducer in the surface of main examination case.No. 1~No. 8 sensors are acceleration transducer, 1
Number~No. 4 sensor arrangements in the radial direction of main examination axle box bearing seat, No. 7 and No. 8 sensor arrangements are in the main axial direction for trying case, No. 5 and 6
Number sensor arrangement is in the radial direction for accompanying examination axle box bearing seat;No. 9 and No. 10 sensors are sonic transducer, be arranged in main examination case with
It accompanies right above examination case at about 40cm;No. 11 sensors are the temperature sensors for testing lubricating oil temperature, are arranged in main examination cabinet,
Lubricating oil temperature is tested in test.Sample frequency 25.6kHz, sampling time 60s, sampling interval 9min, using normal in this test
Method, that is, dead load mode carries out rule in groups, torque 822.7N.M, determines the gear failure when testing gear and broken teeth occurring;
Step 2 carries out feature extraction to the fatigue state of gear, carries out slump evaluations to gear wear degraded performance;
Detachment tooth position in gear fatigue test is selected to set recently and in bearing block location arrangements when carrying out life prediction
463 groups of vibration signals of No. 4 sensors output carry out feature extraction.Test sample frequency 25.6kHz, sampling time 60s, sampling
It is spaced 9min.Sampled point number, which is converted, just can be obtained shown in Fig. 34 by formula (1) after removing singular point for monitoring time
The characteristic value of number sensor output is with monitoring time change curve.
The square amplitude curve of No. 4 sensors is contained from beginning running in stage to completion fatigue test at 77.2h in Fig. 3
The square amplitude of vibration signal that broken teeth occurs, square amplitude y when broken teeth known to gear fatigue test*(T*)=77.375, are denoted as
Degradation, T*For the time for reaching fault threshold for the first time, T*=77.2h;
Step 3, the Real-time Monitoring Data prediction model for establishing the tooth bending fatigue based on LSTM:
Data of the vibration data that acceleration transducer receives after feature extraction are as mode input, the data set
463 group data of the gear from operating normally to breaking down are contained, preceding 300 groups of data is selected to be used for model training, 163 groups of numbers
According to for model measurement.First 300 groups of formula (16) the standardization tooth bending fatigue for training set in application network training
Real-time Monitoring Data, f 'trFor the training set after standardization, it is after dividing that preliminary setting split window length L, which is minimized 2, X,
Mode input, corresponding output is Y, X is inputted hidden layer, by propagated forward (2) formula-(7) formula, obtained output result is
P.To minimize loss function formula (25) as target, learning rate is changed using RMSProp, Nesterov introduces momentum and changes ladder
The method of degree optimizes LSTM model.We (are expressed as application enhancements type LSTM network) predicted.Prediction process is adopted
With the method for iteration.Firstly, the last line data of output Y are as follows:
Yf={ fm-L+1',fm-L+2',…,fm'} (31)
Y is inputtedOutput result can indicate are as follows:
Then the predicted value at m+1 moment is pm+1.Then, by YfRear L-1 data point and pm+1Merge into a new line number
According to:
Yf+1={ fm-L+2',fm-L+3',…,pm+1} (33)
By Yf+1InputThen the predicted value at m+2 moment is pm+2, and so on, obtained forecasting sequence are as follows:
Po={ pm+1,pm+2,…pn} (34)
Next, by P0It carries out z-score instead to standardize and (be expressed as de_zscore), obtain final with test
Collect FteCorresponding forecasting sequence are as follows:
Wherein: m+1≤k≤n, k ∈ N.
Similarly, available with training set F using every a line of X as mode inputtrCorresponding fitting sequence Ptr.Most
Afterwards, by calculating FtrAnd PtrR and FteAnd PteDeviation quantitatively provide fitting and the precision of prediction of model.
Prediction result is as shown in Figure 4.Abscissa is the sample number of acquisition, and ordinate indicates the fitting of the network model and pre-
Survey ability carries out the prediction of degenerate state in 70h, 73.3h, 75h respectively, and the defeated of models fitting is indicated before predicted time point
Data out indicate prediction result after predicted time point.
As frequency of training step=200, the training error loss of model is as shown in figure 5, abscissa is the training time of model
Number, ordinate indicate training error loss.Know the increase with frequency of training, loss is on a declining curve, shows model
Learn the forecasting problem and has predictive ability.LSTM starts to restrain at iteration 300 times or so, and decrease speed is slack-off, explanation
Its error is very small, has good predictive ability.The fitting effect of LSTM model is good, the energy of learning data development law
Power is strong, and it is as shown in table 1 below that sampling number corresponds to the difference between actual value and predicted value.
1 prediction data of table and truthful data compare
Step 4, the parameter based on LSTM prediction model are preferred:
It include many parameters in LSTM prediction model, wherein split window, state vector size is closed the most with learning rate
Key.Then, these hyper parameter cartesian products obtain one group of super parameter.The present invention uses random search, can faster reduce and test
Error is demonstrate,proved, as shown in fig. 6, variation of the discovery data on certain is one-dimensional is more obvious, obtains apparent trend.
Random search programming is simple, using more convenient, can converge to the good value of hyper parameter faster.Grid search makes
With a determining compensation, each step takes a value on regularization coefficient and learning rate.It is different from grid search, it searches at random
Rope does not need the value of discretization hyper parameter, and permission scans in a bigger set, without generating additional calculating
Cost.
The present invention uses random search, is that good hyper parameter can be quickly found because of it, the experiment not wasted, unlike
Grid search provides identical result to two different values of a hyper parameter (it is constant to give other hyper parameter values) sometimes.In net
In lattice search, other hyper parameters will possess identical value in this is tested twice, and in random search, they would generally have
Different values.Therefore, if the verifying collection error corresponding to the variation of the two values does not have significant difference, grid search does not have
It is necessary to repeat two experiments of equal value, and random search still can independently explore other hyper parameters twice.
Step 5 carries out gear predicting residual useful life according to gear condition estimation and known gear distress threshold value:
Data of the vibration data that acceleration transducer receives after feature extraction are as mode input, using improvement
Type LSTM model carries out gear predicting residual useful life.Pass through the degenerate state value predicted and known degenerate state fault threshold
The time for reaching fault threshold for the first time can be solved.Sampling number is converted to the time, absolute error percentage (Absolute is quoted
Percentage Error, APE), TRULFor the remaining life of model prediction.
As can be seen from the above embodiments:
Know that the degenerate state curve predicted in different moments LSTM model is different as shown in Figure 4, it is pre- at 70 hours
Survey state value deviation actual value is more, and error is larger.As monitoring information increases, the degenerate state of LSTM model prediction is connect
Nearly true value, error reduce, can well forecasting system degenerate state mutate after its degenerate state variation and failure
Time of origin.By Fig. 4 (c) it can be seen that when just detecting mutation status point, based on modified LSTM prediction model than very fast
Close to estimative time of day, its dynamic change can be tracked well.
The accurate of remaining life is predicted for a kind of prediction technique of the further real-time remaining life of gear more of the invention
Property, introduce the concept of prediction accuracy, prediction accuracy PA are as follows:
Wherein | TRUL-Ta| indicate absolute value, TRULFor the remaining life of model prediction, TaFor gear real surplus life-span, T*
=77.2h is the gear physical fault time, then Ta=T*-tp, thus can calculate the comparison such as table 2 of LSTM model prediction accuracy
It is shown:
The comparison of 2 model prediction accuracy of table
The invention patent training sample data as seen in Table 2 are more, and predictablity rate is higher.For fatigue accumulation type number
According to long range prediction, LSTM shows clear superiority.
With increasing for monitoring information, prediction accuracy is gradually increased.LSTM prediction model remaining life is gradually at 75h
Close to the real surplus life-span of gear, it can thus be appreciated that the invention patent can be with Accurate Prediction after the mutation of system degradation performance
Gear remaining life, and necessary foundation is provided for adjustment preventive maintenance maintenance policy in time.
In conclusion the present invention is established according to the gear real-time monitoring information received based on LSTM Recognition with Recurrent Neural Network
Fault time forecasting sequence model.Using random search preference pattern hyper parameter, Nesterov momentum optimizes gradient descent algorithm,
Carry out regularized learning algorithm rate using the RMSProp algorithm of Nesterov momentum, by the life search of deep learning frame application to equipment
In, by gear fatigue life verification experimental verification, its prediction model can faster be tracked the dynamic change of system, be improved
Real-time predicting residual useful life accuracy.
The embodiment of the present invention is explained in detail above in conjunction with attached drawing, but the present invention is not limited to above-mentioned implementations
Example, within the knowledge of a person skilled in the art, can also make without departing from the purpose of the present invention
Various change out.
Claims (8)
1. the gear method for predicting residual useful life based on shot and long term memory network, which is characterized in that specific step is as follows:
Step 1 installs sensor in gear-box, obtains the Real-time Monitoring Data of tooth bending fatigue;
Acceleration transducer is mounted on the bearing block position of main examination case, the mounting temperature sensor in gear-box, in main examination case
Noise transducer is installed in surface,
Step 2 carries out feature extraction to the fatigue state of gear, carries out slump evaluations to gear wear degraded performance;
Gear wear condition is changed using square amplitude and carries out performance slump evaluations, in each sampling time Δ t/length,
The square amplitude of the time series of discrete random signal may be expressed as:
Δ t is using the time in formula;N=Fs× Δ t, FsFor sample frequency, n is sampling number, and ∑ indicates to sum, i ∈ (1,2,
3....n), xiIt (t) is sampled value;
Step 3, the Real-time Monitoring Data prediction model for establishing the tooth bending fatigue based on LSTM;
Step 4, the carry out parameter based on LSTM prediction model are preferred;
Step 5 carries out gear predicting residual useful life according to gear condition estimation and known gear distress threshold value:
Data of the vibration data that acceleration transducer receives after feature extraction are as mode input, using modified
LSTM model carries out gear predicting residual useful life, can by the degenerate state value and known degenerate state fault threshold that predict
The time for reaching fault threshold for the first time is solved, sampling number is converted to the time, quotes absolute error percentage (Absolute
Percentage Error,APE)
Wherein, TRULFor the remaining life of model prediction,
Ta=T*-tp
In formula, tpIndicate current time, TaFor the real surplus life-span of gear, T*Reach the time of fault threshold for the first time for gear.
2. the gear method for predicting residual useful life according to claim 1 based on shot and long term memory network, it is characterised in that:
The Real-time Monitoring Data prediction model for the tooth bending fatigue based on LSTM established in step 3 includes input layer, hidden layer, defeated
Five layer, network training and neural network forecast functional modules out;
Input layer is responsible for carrying out preliminary treatment to the Real-time Monitoring Data of tooth bending fatigue to meet network inputs requirement, hides
Layer builds single layer Recognition with Recurrent Neural Network using LSTM cell, and output layer provides prediction result, and network training is used from network structure
Build, learning rate, window setting is upper optimizes to LSTM, and addition Nesterov momentum when parameters revision prevents from falling into office
Portion is optimal, and using the optimization method of RMSProp algorithm correction model adaptive rate, neural network forecast is pre- point by point using the method for iteration
It surveys, and then improves the accuracy of network model prediction.
3. the gear method for predicting residual useful life according to claim 2 based on shot and long term memory network, it is characterised in that:
It includes RNN formula circulation that the LSTM cell that the hidden layer uses, which builds single layer Recognition with Recurrent Neural Network, while in itself " cell "
Portion's circulation, each unit phase input and output, while the door control unit system including more parameters and control information flow, i.e.,
State cell(moment t, cell i), external input door (external input gate) forget door (forget gate),
Out gate (output gate);
Forget door to be responsible for removing the unessential information of LSTM study from " cell ", these information will be moved by door control unit operation
It removes, forgets door and take two input ht-1And xt。ht-1For the state value of previous unit, xtFor the value of the t moment of list entries x, add
Add bias term to input Sigmoid function.Sigmoid function value range is [0,1], if value is zero, forgets door and allows for
Location mode forgets the information;
Corresponding propagated forward algorithm is as follows:
Input gate:
Forget door:
Cell:
Out gate:
Cell output:
Backpropagation calculates as follows:
Cell output:
Out gate:
State:
Cell:
Forget door:
Input gate:
Wherein ∑ indicates summation,Current input vector, subscript l, φ, c, ω respectively indicate indicate input gate, forget door,
The relevant parameter of cell state unit, out gate.wil、wiφ、wicAnd wiwIt respectively indicates from being input to input gate, forget door, thin
The weight of born of the same parents' state cell, out gate, whl、whφ、whcAnd whwIt respectively indicates from hidden layer to input gate, forget door, cell state
The weight of unit, out gate, wcl、wcφAnd wcwIt respectively indicates from cell state to input gate, the weight of forgetting door and out gate,
atIndicate the input of t moment, btIndicate the output of t moment,Indicate the state of t moment cell, g and h respectively indicate the defeated of cell
Enter and export activation primitive, I indicates the number of the neuron of input layer, and K indicates that the neuron number of output layer, H indicate hidden layer
Number of cells, σ are sigmoid functions.Indicate the cell output of t moment,Indicate the state of t moment, ζ indicates loss letter
Number, δtIndicate gradient error, wckAnd wcgIt respectively indicates from cell state to output layer and the weight of hidden layer.
4. the gear method for predicting residual useful life according to claim 2 based on shot and long term memory network, it is characterised in that:
The method of the network training is as follows:
First in input layer, the Real-time Monitoring Data for defining tooth bending fatigue is F0={ f1,f2,…,fn, then the instruction divided
F can be expressed as by practicing collection and test settr={ f1,f2,…,fmAnd Fte={ fm+1,fm+2,…,fn, meet constraint condition m < n
And m, n ∈;
Then to the element f in training settIt is standardized, uses classical z-score standardization formula (mean value 0, standard
Difference is 1, is expressed as zscore), the training set after standardization can indicate are as follows:
Ftr'={ f1',f2',…,fm'}
Meet:
1≤t≤m,t∈
The characteristics of in order to adapt to hidden layer input, the method f ' divided using datatrIt is handled, setting split window length takes
Value is L, then the mode input after dividing are as follows:
X={ X1,X2,…,XL}
Xp={ fp',fp+1',…,fm-L+p-1'}
Meet:
1≤p≤L;p,L∈N
Corresponding output are as follows:
Y={ Y1,Y2,…YL}
Yp={ fp+1',fp+2',…,fm-L+p'}
Next, X is inputted hidden layer, since hidden layer includes the L isomorphism LSTM cells by the connection of front and back moment, process is hidden
Output after hiding layer can indicate are as follows:
P={ P1,P2,…,PL}
Pp=LSTMforward(Xp,Cp-1,Hp-1)
Wherein, Cp-1And Hp-1The state and output of respectively previous LSTM cell, LSTMforwardTo meter before expression LSTM cell
Calculation method sets cell state vector magnitude as Sstate, then Cp-1And Hp-12 vector magnitudes are Sstate, hidden layer P, model
Input X and theoretical output Y is the two-dimensional array that dimension is (m-L, L), selects mean square error as error calculation formula, training
The loss function of process can be with is defined as:
The minimum optimization aim of loss function is set, the train epochs steps of netinit is given, changes using RMSProp
Learning rate, Nesterov introduce momentum and change gradient, change update mode in terms of two, and then obtain final hidden layer net
Network;
Finally, optimizing loss function using stochastic gradient descent, stochastic gradient descent is small quantities of according to data generation distribution extraction m
(independent identically distributed) sample is measured, iteration only calculates the loss of a sample each time, then traverses all samples again, completes one
The calculating of wheel.By calculating their gradient mean value, the unbiased esti-mator of available gradient is expressed as follows
Wherein g is gradient, and θ is the initial parameter of LSTM model forward calculation,It is gradient element, indicatesLocal derviation is asked to θ, using the sample from small lot m, stochastic gradient descent algorithm is declined using following gradient
Estimation
θ←θ-ηg
Wherein η is learning rate, and L is the loss of each sample
L (x, y, θ)=- logp (y | x;θ)
Wherein, logarithm is sought in log expression.
5. the gear method for predicting residual useful life according to claim 2 based on shot and long term memory network, it is characterised in that:
Specific step is as follows for the Nesterov momentum parameter optimization:
Assuming that learning rate η, momentum parameter α, initial rate v
It is iterated process every time:
1. extracting the small lot sample of { x (1), x (2) ... ..x (m) } from training set, wherein xtCorresponding output yt,
2. calculate gradient and Loss, v and θ are updated with this:
Calculate gradient:
It is interim to update:
θ←θ+v。
6. the gear method for predicting residual useful life according to claim 2 based on shot and long term memory network, it is characterised in that:
Specific step is as follows for the RMSProp variable learning rate arithmetic: assuming that learning rate η, momentum parameter α, initial rate v,
Rate of decay ρ initializes cumulant variable γ=0
It is iterated process every time:
1. extracting the small lot sample of { x (1), x (2) ... ..x (m) } from training set, wherein xtCorresponding output yt,
2. calculating gradient and Loss, γ is updated, calculates and updates in conjunction with gradient:
Calculate gradient:
It is interim to update:
θ←θ+v。
7. the gear method for predicting residual useful life according to claim 2 based on shot and long term memory network, it is characterised in that:
Modified LSTM network (is expressed as) prediction process are as follows:
Firstly, the last line data of output Y are as follows:
Yf={ fm-L+1',fm-L+2',…,fm'}
Y is inputtedOutput result can indicate are as follows:
Then the predicted value at m+1 moment is pm+1.Then, by YfRear L-1 data point and pm+1Merge into new data line:
Yf+1={ fm-L+2',fm-L+3',…,pm+1}
By Yf+1InputThen the predicted value at m+2 moment is pm+2, and so on, obtained forecasting sequence are as follows:
Po={ pm+1,pm+2,…pn}
Next, by P0It carries out z-score instead to standardize and (be expressed as de_zscore), obtain final with test set Fte
Corresponding forecasting sequence are as follows:
Wherein: m+1≤k≤n, k ∈ N.
Similarly, available with training set F using every a line of X as mode inputtrCorresponding fitting sequence Ptr.Finally, logical
Cross calculating FtrAnd PtrAnd FteAnd PteDeviation quantitatively provide fitting and the precision of prediction of model.
8. the gear method for predicting residual useful life according to claim 1 based on shot and long term memory network, it is characterised in that:
It is preferred to the carry out parameter based on LSTM prediction model by the way of random search in the step 3.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811096152.0A CN109343505A (en) | 2018-09-19 | 2018-09-19 | Gear method for predicting residual useful life based on shot and long term memory network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811096152.0A CN109343505A (en) | 2018-09-19 | 2018-09-19 | Gear method for predicting residual useful life based on shot and long term memory network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109343505A true CN109343505A (en) | 2019-02-15 |
Family
ID=65305924
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811096152.0A Pending CN109343505A (en) | 2018-09-19 | 2018-09-19 | Gear method for predicting residual useful life based on shot and long term memory network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109343505A (en) |
Cited By (33)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110175425A (en) * | 2019-05-31 | 2019-08-27 | 重庆大学 | A kind of prediction technique of the gear remaining life based on MMALSTM |
CN110210126A (en) * | 2019-05-31 | 2019-09-06 | 重庆大学 | A kind of prediction technique of the gear remaining life based on LSTMPP |
CN110276441A (en) * | 2019-05-20 | 2019-09-24 | 成都理工大学 | A kind of trapezoidal overlap kernel impulse response estimation method based on deep learning |
CN110366232A (en) * | 2019-06-19 | 2019-10-22 | 东南大学 | Sensor transmissions energy control method for remote status estimation |
CN110378870A (en) * | 2019-06-06 | 2019-10-25 | 西安交通大学 | A kind of turbine blade erosion degree method of discrimination based on ResNet-GRU network |
CN110543869A (en) * | 2019-09-10 | 2019-12-06 | 哈工大机器人(山东)智能装备研究院 | Ball screw service life prediction method and device, computer equipment and storage medium |
CN110552804A (en) * | 2019-09-05 | 2019-12-10 | 重庆邮电大学 | method for estimating combustion timing of mixed fuel HCCI (homogeneous charge compression ignition) engine based on LSTM (localized fuel transfer model) |
CN110909509A (en) * | 2019-11-28 | 2020-03-24 | 哈尔滨理工大学 | Bearing life prediction method based on InfoLSGAN and AC algorithm |
CN110956112A (en) * | 2019-11-25 | 2020-04-03 | 南京工业大学 | Novel high-reliability slewing bearing life evaluation method |
CN111475987A (en) * | 2020-04-02 | 2020-07-31 | 重庆大学 | Gear residual life prediction method based ON SAE and ON-L STM |
CN111476400A (en) * | 2020-03-11 | 2020-07-31 | 珠海格力电器股份有限公司 | Circuit fault prediction method, device, equipment and computer readable medium |
CN111475986A (en) * | 2020-04-02 | 2020-07-31 | 重庆大学 | L STM-AON-based gear residual life prediction method |
CN111475948A (en) * | 2020-04-07 | 2020-07-31 | 青岛理工大学 | Novel prediction method for settlement during subway tunnel construction based on L STM network |
CN111475921A (en) * | 2020-03-13 | 2020-07-31 | 重庆邮电大学 | Tool residual life prediction method based on edge calculation and L STM network |
CN111537888A (en) * | 2020-05-09 | 2020-08-14 | 国网福建省电力有限公司莆田供电公司 | Data-driven echelon battery SOC prediction method |
CN111948563A (en) * | 2020-06-19 | 2020-11-17 | 浙江大学 | Electric forklift lithium battery residual life prediction method based on multi-neural network coupling |
CN112001113A (en) * | 2020-07-02 | 2020-11-27 | 浙江大学 | Battery life prediction method based on particle swarm optimization long-time memory network |
CN112101521A (en) * | 2020-08-13 | 2020-12-18 | 国网辽宁省电力有限公司电力科学研究院 | Building energy consumption prediction method based on long-term and short-term memory network hybrid model |
CN112214852A (en) * | 2020-10-09 | 2021-01-12 | 电子科技大学 | Degradation rate-considered turbine mechanical performance degradation prediction method |
CN112235043A (en) * | 2020-09-14 | 2021-01-15 | 上海大学 | Distributed optical fiber abnormal data restoration model based on self-adaptive long-term and short-term memory |
CN112395806A (en) * | 2020-11-11 | 2021-02-23 | 北京京航计算通讯研究所 | Method and device for predicting residual life of comprehensive transmission hydraulic system |
CN112446537A (en) * | 2020-11-20 | 2021-03-05 | 国网浙江省电力有限公司宁波供电公司 | Short-term load prediction method based on deep long-term and short-term memory network |
CN112580263A (en) * | 2020-12-24 | 2021-03-30 | 湖南工业大学 | Turbofan engine residual service life prediction method based on space-time feature fusion |
CN112782498A (en) * | 2019-11-11 | 2021-05-11 | 株洲中车时代电气股份有限公司 | Fault monitoring method and device for capacitor |
CN112800141A (en) * | 2020-12-11 | 2021-05-14 | 广东海洋大学 | On-demand service aggregation and recommendation method based on RGPS meta-model |
CN113256865A (en) * | 2020-11-06 | 2021-08-13 | 上海兴容信息技术有限公司 | Control method and system of intelligent access control |
CN113821974A (en) * | 2021-09-07 | 2021-12-21 | 北京大学 | Engine residual life prediction method based on multiple failure modes |
CN114008549A (en) * | 2019-04-23 | 2022-02-01 | 大众汽车股份公司 | Method for determining remaining usage period, remaining usage period determining circuit, remaining usage period determining apparatus |
CN117030268A (en) * | 2023-10-07 | 2023-11-10 | 太原科技大学 | Rolling bearing fault diagnosis method |
CN117350174A (en) * | 2023-12-04 | 2024-01-05 | 国网天津市电力公司营销服务中心 | Method, system, electronic equipment and medium for predicting residual life of intelligent ammeter |
WO2024021108A1 (en) * | 2022-07-29 | 2024-02-01 | Siemens Aktiengesellschaft | Method and device for predicting service life of rolling bearing and computer readable storage medium |
CN117875191A (en) * | 2024-03-08 | 2024-04-12 | 东莞市星火齿轮有限公司 | Gearbox service life assessment method and system based on big data and storage medium |
CN111475987B (en) * | 2020-04-02 | 2024-05-24 | 重庆大学 | SAE and ON-LSTM-based gear residual life prediction method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105740625A (en) * | 2016-01-31 | 2016-07-06 | 太原科技大学 | Real time residual life prediction method of gear |
CN108304927A (en) * | 2018-01-25 | 2018-07-20 | 清华大学 | Bearing fault modality diagnostic method and system based on deep learning |
CN108520320A (en) * | 2018-03-30 | 2018-09-11 | 华中科技大学 | A kind of equipment life prediction technique based on multiple shot and long term memory network and Empirical Bayes |
-
2018
- 2018-09-19 CN CN201811096152.0A patent/CN109343505A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105740625A (en) * | 2016-01-31 | 2016-07-06 | 太原科技大学 | Real time residual life prediction method of gear |
CN108304927A (en) * | 2018-01-25 | 2018-07-20 | 清华大学 | Bearing fault modality diagnostic method and system based on deep learning |
CN108520320A (en) * | 2018-03-30 | 2018-09-11 | 华中科技大学 | A kind of equipment life prediction technique based on multiple shot and long term memory network and Empirical Bayes |
Non-Patent Citations (2)
Title |
---|
BVL10101111: "《CSDN博客》", 21 May 2017 * |
王鑫等: "基于LSTM循环神经网络的故障时间序列预测", 《北京航空航天大学学报》 * |
Cited By (47)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114008549A (en) * | 2019-04-23 | 2022-02-01 | 大众汽车股份公司 | Method for determining remaining usage period, remaining usage period determining circuit, remaining usage period determining apparatus |
CN110276441B (en) * | 2019-05-20 | 2021-05-04 | 成都理工大学 | Trapezoidal overlapped kernel pulse estimation method based on deep learning |
CN110276441A (en) * | 2019-05-20 | 2019-09-24 | 成都理工大学 | A kind of trapezoidal overlap kernel impulse response estimation method based on deep learning |
CN110210126A (en) * | 2019-05-31 | 2019-09-06 | 重庆大学 | A kind of prediction technique of the gear remaining life based on LSTMPP |
CN110175425B (en) * | 2019-05-31 | 2023-02-21 | 重庆大学 | Prediction method of residual life of gear based on MMALSTM |
CN110175425A (en) * | 2019-05-31 | 2019-08-27 | 重庆大学 | A kind of prediction technique of the gear remaining life based on MMALSTM |
CN110378870A (en) * | 2019-06-06 | 2019-10-25 | 西安交通大学 | A kind of turbine blade erosion degree method of discrimination based on ResNet-GRU network |
CN110378870B (en) * | 2019-06-06 | 2021-04-20 | 西安交通大学 | ResNet-GRU network-based turbine blade water erosion degree judging method |
CN110366232A (en) * | 2019-06-19 | 2019-10-22 | 东南大学 | Sensor transmissions energy control method for remote status estimation |
CN110366232B (en) * | 2019-06-19 | 2022-02-11 | 东南大学 | Sensor transmission energy control method for remote state estimation |
CN110552804A (en) * | 2019-09-05 | 2019-12-10 | 重庆邮电大学 | method for estimating combustion timing of mixed fuel HCCI (homogeneous charge compression ignition) engine based on LSTM (localized fuel transfer model) |
CN110543869A (en) * | 2019-09-10 | 2019-12-06 | 哈工大机器人(山东)智能装备研究院 | Ball screw service life prediction method and device, computer equipment and storage medium |
CN112782498A (en) * | 2019-11-11 | 2021-05-11 | 株洲中车时代电气股份有限公司 | Fault monitoring method and device for capacitor |
CN112782498B (en) * | 2019-11-11 | 2022-07-29 | 株洲中车时代电气股份有限公司 | Fault monitoring method and device for capacitor |
CN110956112B (en) * | 2019-11-25 | 2023-07-18 | 南京工业大学 | Novel high-reliability slewing bearing service life assessment method |
CN110956112A (en) * | 2019-11-25 | 2020-04-03 | 南京工业大学 | Novel high-reliability slewing bearing life evaluation method |
CN110909509A (en) * | 2019-11-28 | 2020-03-24 | 哈尔滨理工大学 | Bearing life prediction method based on InfoLSGAN and AC algorithm |
CN111476400A (en) * | 2020-03-11 | 2020-07-31 | 珠海格力电器股份有限公司 | Circuit fault prediction method, device, equipment and computer readable medium |
CN111475921A (en) * | 2020-03-13 | 2020-07-31 | 重庆邮电大学 | Tool residual life prediction method based on edge calculation and L STM network |
CN111475986B (en) * | 2020-04-02 | 2024-05-24 | 重庆大学 | LSTM-AON-based gear residual life prediction method |
CN111475987B (en) * | 2020-04-02 | 2024-05-24 | 重庆大学 | SAE and ON-LSTM-based gear residual life prediction method |
CN111475986A (en) * | 2020-04-02 | 2020-07-31 | 重庆大学 | L STM-AON-based gear residual life prediction method |
CN111475987A (en) * | 2020-04-02 | 2020-07-31 | 重庆大学 | Gear residual life prediction method based ON SAE and ON-L STM |
CN111475948A (en) * | 2020-04-07 | 2020-07-31 | 青岛理工大学 | Novel prediction method for settlement during subway tunnel construction based on L STM network |
CN111537888A (en) * | 2020-05-09 | 2020-08-14 | 国网福建省电力有限公司莆田供电公司 | Data-driven echelon battery SOC prediction method |
CN111948563A (en) * | 2020-06-19 | 2020-11-17 | 浙江大学 | Electric forklift lithium battery residual life prediction method based on multi-neural network coupling |
CN112001113B (en) * | 2020-07-02 | 2023-12-19 | 浙江大学 | Battery life prediction method based on particle swarm optimization long-time and short-time memory network |
CN112001113A (en) * | 2020-07-02 | 2020-11-27 | 浙江大学 | Battery life prediction method based on particle swarm optimization long-time memory network |
CN112101521A (en) * | 2020-08-13 | 2020-12-18 | 国网辽宁省电力有限公司电力科学研究院 | Building energy consumption prediction method based on long-term and short-term memory network hybrid model |
CN112235043B (en) * | 2020-09-14 | 2022-12-23 | 上海大学 | Distributed optical fiber abnormal data restoration device based on self-adaptive long-term and short-term memory |
CN112235043A (en) * | 2020-09-14 | 2021-01-15 | 上海大学 | Distributed optical fiber abnormal data restoration model based on self-adaptive long-term and short-term memory |
CN112214852A (en) * | 2020-10-09 | 2021-01-12 | 电子科技大学 | Degradation rate-considered turbine mechanical performance degradation prediction method |
CN113256865A (en) * | 2020-11-06 | 2021-08-13 | 上海兴容信息技术有限公司 | Control method and system of intelligent access control |
CN112395806A (en) * | 2020-11-11 | 2021-02-23 | 北京京航计算通讯研究所 | Method and device for predicting residual life of comprehensive transmission hydraulic system |
CN112395806B (en) * | 2020-11-11 | 2021-11-09 | 北京京航计算通讯研究所 | Method and device for predicting residual life of comprehensive transmission hydraulic system |
CN112446537A (en) * | 2020-11-20 | 2021-03-05 | 国网浙江省电力有限公司宁波供电公司 | Short-term load prediction method based on deep long-term and short-term memory network |
CN112800141A (en) * | 2020-12-11 | 2021-05-14 | 广东海洋大学 | On-demand service aggregation and recommendation method based on RGPS meta-model |
CN112580263A (en) * | 2020-12-24 | 2021-03-30 | 湖南工业大学 | Turbofan engine residual service life prediction method based on space-time feature fusion |
CN113821974B (en) * | 2021-09-07 | 2023-11-24 | 北京大学 | Engine residual life prediction method based on multiple fault modes |
CN113821974A (en) * | 2021-09-07 | 2021-12-21 | 北京大学 | Engine residual life prediction method based on multiple failure modes |
WO2024021108A1 (en) * | 2022-07-29 | 2024-02-01 | Siemens Aktiengesellschaft | Method and device for predicting service life of rolling bearing and computer readable storage medium |
CN117030268B (en) * | 2023-10-07 | 2024-01-23 | 太原科技大学 | Rolling bearing fault diagnosis method |
CN117030268A (en) * | 2023-10-07 | 2023-11-10 | 太原科技大学 | Rolling bearing fault diagnosis method |
CN117350174A (en) * | 2023-12-04 | 2024-01-05 | 国网天津市电力公司营销服务中心 | Method, system, electronic equipment and medium for predicting residual life of intelligent ammeter |
CN117350174B (en) * | 2023-12-04 | 2024-04-02 | 国网天津市电力公司营销服务中心 | Method, system, electronic equipment and medium for predicting residual life of intelligent ammeter |
CN117875191A (en) * | 2024-03-08 | 2024-04-12 | 东莞市星火齿轮有限公司 | Gearbox service life assessment method and system based on big data and storage medium |
CN117875191B (en) * | 2024-03-08 | 2024-05-28 | 东莞市星火齿轮有限公司 | Gearbox service life assessment method and system based on big data and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109343505A (en) | Gear method for predicting residual useful life based on shot and long term memory network | |
CN105973594B (en) | A kind of rolling bearing fault Forecasting Methodology based on continuous depth confidence network | |
CN109885907B (en) | Cloud model-based satellite attitude control system health state assessment and prediction method | |
CN109063308B (en) | Health assessment method based on deep quantum learning | |
Yan et al. | A dynamic multi-scale Markov model based methodology for remaining life prediction | |
CN108596327B (en) | Seismic velocity spectrum artificial intelligence picking method based on deep learning | |
CN112418406A (en) | Wind power tower inclination angle missing data supplementing method based on SSA-LSTM model | |
CN111581880A (en) | Rolling bearing residual life prediction method based on APSO-DBN | |
CN116010900A (en) | Multi-scale feature fusion gearbox fault diagnosis method based on self-attention mechanism | |
CN109447305B (en) | Trend prediction method based on quantum weighted long-time and short-time memory neural network | |
CN114707712A (en) | Method for predicting requirement of generator set spare parts | |
Sahu et al. | Use of ANN and neuro fuzzy model to predict bearing capacity factor of strip footing resting on reinforced sand and subjected to inclined loading | |
Göktepe et al. | Comparison of multilayer perceptron and adaptive neuro-fuzzy system on backcalculating the mechanical properties of flexible pavements | |
CN110414666A (en) | Pumping plant unit fault recognition method based on improved LSTM network | |
CN111191823B (en) | Deep learning-based production logistics prediction method | |
CN115795343A (en) | Method for analyzing state of unbalanced sample ammeter production equipment based on digital twin | |
Sun et al. | Pruning Elman neural network and its application in bolt defects classification | |
CN111723523A (en) | Estuary residual water level prediction method based on cascade neural network | |
CN114166509A (en) | Motor bearing fault prediction method | |
Ma et al. | Rolling-optimized model predictive vibration controller for offshore platforms subjected to random waves and winds under uncertain sensing delay | |
CN116303786B (en) | Block chain financial big data management system based on multidimensional data fusion algorithm | |
CN111241749B (en) | Permanent magnet synchronous motor chaos prediction method based on reserve pool calculation | |
CN112132324A (en) | Ultrasonic water meter data restoration method based on deep learning model | |
Sulistiyo et al. | Evolution strategies for weight optimization of Artificial Neural Network in time series prediction | |
CN111414927A (en) | Method for evaluating seawater quality |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190215 |